一种新的多类模式识别方法  

NOVEL APPROACH TO MULTI-CLASS CLASSIFICATION

作　　者：方勇[1] 戚飞虎[1] 

机构地区：[1]上海交通大学计算机科学与工程系,上海200030

出　　处：《红外与毫米波学报》2004年第6期418-422,共5页Journal of Infrared and Millimeter Waves

基　　金：国家自然科学基金资助项目(60072029;60271033)

摘　　要：提出了一种基于支撑向量机的多类分类器,用N1个支撑向量机组合构成一个具有二叉树结构形式的N多类分类器.讨论了该多类分类器的泛化推广能力,同时还提出了该多类分类器的基于特征空间的BTSVM学习算法.BTSVM算法使用核函数转换的方式计算特征空间的样本距离;采用类间最小距离最大化作为聚类准则,在每个决策结点产生两个最优子集;然后采用支撑向量机学习算法学习两个最优子集,确定决策结点的最优分类面.理论和实验结果表明,本文提出的基于支撑向量机的多类分类器在整体性能上要优于其它类似的分类器系统.A new multi-classifier was proposed based on support vector machines for a N-class classification problem, which comprised N-1 support vector machines in the form of a binary tree. The generalization performance of multi-classifiers was discussed, and a new learning algorithm, the BTSVM algorithm, was presented based on high-dimension feature spaces. The BTSVM algorithm evaluates example distances by kernel functions, employs the maximization of minimum distances as clustering criteria to obtain two optimal subsets, and generates the optimal classification functions with support vector machine leaning algorithms at each decision node. Theoretical analysis and experimental results show that the BTSVM algorithm is superior to other competitive multi-classifiers.

关 键 词：支撑向量机 多类分类 多类模式识别 学习算法 基于特征 分类器 结点 类似 最优子集 核函数 

分 类 号：TN911[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]